Serving a Million Models： How Hugging Face Scaled with MongoDB Atlas ｜ MongoDB.local London 2026

In this session, explore how Hugging Face, the world’s largest open source AI platform, scales to support over 13 million users, more than 2 million public models, and 500,000+ datasets, used by 50,000+ organizations, including over 30% of the Fortune 500. You’ll learn how the team leverages MongoDB Atlas to maintain low latency and a streamlined operational footprint as usage grows globally, along with key architecture decisions, trade offs, and lessons learned from building at scale.

Through a live demo of the Hugging Face Hub, Arek Borucki, Machine Learning Platform & Database Engineer at Hugging Face will bring these concepts to life by showcasing models, datasets, and platform functionality in action, while highlighting how MongoDB powers the platform behind the scenes to enable performance, scalability, and a seamless developer experience.

00:00:00 - Introduction & Who is Hugging Face?
00:01:51 - Understanding the Scale: Millions of Users, Models & Datasets
00:04:17 - High-Level Architecture: How Hugging Face Uses MongoDB Atlas
00:06:05 - Live Demo: Models, Datasets, and Spaces Pillars
00:11:19 - Introducing a New Feature: Storage Buckets
00:13:19 - Interacting with the Hub via Hugging Face CLI
00:14:55 - Infrastructure Breakdown: 7-Node Replica Set Cluster
00:19:47 - Offloading Traffic: Read Preferences & Hidden Analytics Nodes
00:21:08 - Data Modeling: Compute Patterns & Document Embedding Limits
00:22:43 - Index Management & Atlas Search (Full-Text Search)
00:25:19 - The Move to Sharding: Choosing Shard Keys with Analyzers
00:27:30 - Hugging Chat Database Architecture
00:28:30 - Mongoku: Open-Source Web GUI for MongoDB